Google Gemini-Powered PageBreak Agent Finds More Than 500 XSS Vulnerabilities

Google’s Product Security team has developed PageBreak, a Gemini-powered security agent that has found more than 500 cross-site scripting (XSS) vulnerabilities across Google’s first-party web applications since November 2025, according to the supplied project summary.

The agent mainly uses Gemini 3.1 Pro and Gemini 3.5 Flash. Its central feature is deterministic validation, instead of reporting every suspicious code pattern, it tests potential vulnerabilities against running applications and checks whether exploitation actually works. Google says this approach delivers a near-zero false positive rate.

Google security engineer Michał Bentkowski described three high-severity findings in a technical report published on September 24, 2026.

These included cache poisoning, an administrative console XSS flaw, and a browser extension vulnerability. Google security engineers and external researchers had previously examined the affected applications without identifying these issues.

Gemini Finds 500+ XSS

PageBreak separates vulnerability reasoning from proof. The agent identifies a possible weakness, then a validator executes a real payload to determine whether the suspected attack succeeds.

For XSS, that means confirming JavaScript execution rather than relying only on generated explanations or source-code patterns.

According to the supplied summary, validators can also test path traversal by writing a file to a world-readable location, or check remote code execution through an outbound DNS request. Unconfirmed findings become starting points for deeper scans and help engineers develop additional validation tools.

Google’s engineering environment supports this process. Its shared code repository lets PageBreak follow execution paths across applications. Live HTTP traffic connects requests to source code, while an existing authenticated scanner provides access to many internal services.

The agent also tested applications built with Google’s high-assurance web framework. Across hundreds of these applications, the supplied summary reports only two XSS findings, both involving internal applications or debug endpoints with hardening gaps.

In apis.google.com, PageBreak discovered a cache poisoning flaw involving an unchecked URL path segment. The server inserted that segment into returned JavaScript but excluded it from the cache key.

An attacker could therefore make malicious and legitimate requests share a cached response, causing injected code to reach other users.

Google said the poisoned cache affected users near the attack’s geographical location rather than everyone globally.

However, the flaw could have affected sensitive Google services and external websites loading the JavaScript. Google found no evidence of real-world exploitation of this vulnerability.

In admin.google.com, an unchecked redirect parameter allowed a JavaScript URL to reach window.location.

Although a cryptographic signature protected the request, PageBreak discovered that another authorization endpoint could generate a valid signature for the malicious input, completing the attack chain, tldrsec said.

The third finding affected the Tag Assistant Extension. Weak origin checks, a recoverable nonce, and unsafe message forwarding allowed attacker-controlled script content to reach a debugged page. Acceptance of data URLs enabled arbitrary JavaScript execution, creating a universal XSS vulnerability.

Join 16,000+ SOC teams using ANY.RUN to streamline threat investigations and reduce manual effort. Explore for your team 

Varshini
Varshini
Varshini is a Cyber Security expert in Threat Analysis, Vulnerability Assessment, and Research. Passionate about staying ahead of emerging Threats and Technologies..

Trending News

Related Stories